Atomic wavefunction initialization in ab initio molecular dynamics using distributed Lanczos

We present a distributed scheme for initialization from atomic wavefunctions in ab initio molecular dynamics simulations. Good initial guesses for approximate wavefunctions are very important in order to enable practical simulations with thousands of atoms. The new scheme is based on a distributed implementation of the Lanczos algorithm for very large dense eigenproblems. We show that the massively parallel BG/L (Blue Gene/L) supercomputer with its very fast separate network for collective communications is an ideal platform for the parallel Lanczos algorithm. We have implemented the new scheme in the popular plane-wave code CPMD. We showcase the applicability of the distributed initialization by a series of examples on a family of Silicon super cells ranging from 512 to 2048 atoms.